AI-linked pay needs an opportunity-adjusted evidence rubric

Selby Jennings’ AI Talent Pulse report, covered in reporting on 14 October, says 25% of respondents say their organisation ties raises, promotions or bonuses to AI proficiency. Thirty percent also report that their organisation requires disclosure of AI use in formal performance reviews. The primary methodology identifies 500 US HR professionals in finance, aged 22–64, surveyed by Censuswide on 19–26 August 2026.

That denominator matters. Some news coverage describes executives or firms; the primary report describes respondents. It does not provide a census of financial institutions or demonstrate that an AI-linked pay policy improves performance. The public page reports 42% with new AI-skill assessment guidelines and 30% with new roles or departments dedicated to proving AI returns. These remain self-reported practices.

Decide what proficiency means before attaching money

The managerial question is no longer simply whether people should learn AI. It is what evidence can legitimately influence a consequential employment decision. A prompt count, licence activation or self-reported use can show activity without showing sound judgement. Conversely, a worker may demonstrate capability by declining to use AI where the task is unsuitable.

Start with a role-specific task and a visible quality standard. Ask what a competent person must produce, which errors matter, what independent verification is required and when escalation is the correct result. An employee should be able to explain the decision, not merely display a polished output. This is a proposed assessment design, not a validated rubric supplied by the survey.

Separate three records: opportunity to practise, demonstrated capability and operating outcome. Opportunity includes access to approved tools, training time, suitable tasks and necessary data. Capability includes checking, correcting and safely declining a tool. Outcome concerns accepted work after verification, including rework and supervisory effort. Keeping them distinct avoids treating an access constraint as proof of weak performance.

The earlier newsroom discussion of practice access behind the “superuser” label is relevant here. It does not prove a pay disparity in this sample; it identifies a question an employer should investigate before comparing employees.

Make the evidence contestable

A pilot could use a small, representative task set with an agreed assistance policy. Give participants comparable tools and practice time, preserve the task context, and have reviewers score against a written rubric without knowing whose output they are judging where feasible. Include exceptions, corrections and unaided explanation rather than rewarding speed alone.

Do not move directly from a pilot score to pay. First inspect whether results change across tasks, reviewers and tool versions. Record missing access, unsuitable work assignments and disagreements. Provide a route to correct a mistaken record and show what evidence changed a decision. Required employment and domain review remain pending; this article is not advice on the legality of a compensation scheme.

NIST’s generative-AI risk profile recommends context-appropriate empirical evaluation and cautions against extrapolating from narrow assessments. It does not certify an employee rubric or corroborate this survey. Its useful contribution is the discipline of testing the proposed measurement before acting on it.

The survey is a signal that AI proficiency is entering consequential management systems, not proof that a particular pay formula works. Before adding another rating field, require a documented task standard, comparable opportunity and a correction route. Reward verified contribution and judgement; keep tool activity as supporting context rather than the verdict.